Re: [PATCH 00/23] kconfig: move compiler capability tests to Kconfig

[Masahiro Yamada]

2018-02-22 6:39 GMT+09:00 Ulf Magnusson <ulfalizer@xxxxxxxxx>:
> On Wed, Feb 21, 2018 at 09:57:03PM +0900, Masahiro Yamada wrote:
>> 2018-02-21 19:52 GMT+09:00 Arnd Bergmann <arnd@xxxxxxxx>:
>> > On Wed, Feb 21, 2018 at 11:20 AM, Masahiro Yamada
>> > <yamada.masahiro@xxxxxxxxxxxxx> wrote:
>> >> 2018-02-21 18:56 GMT+09:00 Arnd Bergmann <arnd@xxxxxxxx>:
>> >>> On Wed, Feb 21, 2018 at 8:38 AM, Masahiro Yamada
>> >>> <yamada.masahiro@xxxxxxxxxxxxx> wrote:
>> >>>> 2018-02-20 0:18 GMT+09:00 Ulf Magnusson <ulfalizer@xxxxxxxxx>:
>> >>
>> >> Let me clarify my concern.
>> >>
>> >> When we test the compiler flag, is there a case
>> >> where a particular flag depends on -m{32,64} ?
>> >>
>> >> For example, is there a compiler that supports -fstack-protector
>> >> for 64bit mode, but unsupports it for 32bit mode?
>> >>
>> >>   $(cc-option -m32)                     ->  y
>> >>   $(cc-option -m64)                     ->  y
>> >>   $(cc-option -fstack-protector)        ->  y
>> >>   $(cc-option -m32 -fstack-protector)   ->  n
>> >>   $(cc-option -m64 -fstack-protector)   ->  y
>> >>
>> >> I guess this is unlikely to happen,
>> >> but I am not whether it is zero possibility.
>> >>
>> >> If this could happen,
>> >> $(cc-option ) must be evaluated together with
>> >> correct bi-arch option (either -m32 or -m64).
>> >>
>> >>
>> >> Currently, -m32/-m64 is specified in Makefile,
>> >> but we are moving compiler tests to Kconfig
>> >> and, CONFIG_64BIT can be dynamically toggled in Kconfig.
>> >
>> > I don't think it can happen for this particular combination (stack protector
>> > and word size), but I'm sure we'll eventually run into options that
>> > need to be tested in combination. For the current CFLAGS_KERNEL
>> > setting, we definitely have the case of needing the variables to be
>> > evaluated in a specific order.
>> >
>>
>>
>>
>>
>> I was thinking of how we can handle complex cases
>> in the current approach.
>>
>>
>>
>> (Case 1)
>>
>> Compiler flag -foo and -bar interacts, so
>> we also need to check the combination of the two.
>>
>>
>> config CC_HAS_FOO
>>         def_bool $(cc-option -foo)
>>
>> config CC_HAS_BAR
>>         def_bool $(cc-option -bar)
>>
>> config CC_HAS_FOO_WITH_BAR
>>         def_bool $(cc-option -foo -bar)
>>
>>
>>
>> (Case 2)
>> Compiler flag -foo is sensitive to word-size.
>> So, we need to test this option together with -m32/-m64.
>> User can toggle CONFIG_64BIT, like i386/x86_64.
>>
>>
>> config CC_NEEDS_M64
>>           def_bool $(cc-option -m64) && 64BIT
>>
>> config CC_NEEDS_M32
>>           def_bool $(cc-option -m32) && !64BIT
>>
>> config CC_HAS_FOO
>>          bool
>>          default $(cc-option -m64 -foo) if CC_NEEDS_M64
>>          default $(cc-option -m32 -foo) if CC_NEEDS_M32
>>          default $(cc-option -foo)
>>
>>
>>
>> (Case 3)
>> Compiler flag -foo is sensitive to endian-ness.
>>
>>
>> config CC_NEEDS_BIG_ENDIAN
>>           def_bool $(cc-option -mbig-endian) && CPU_BIG_ENDIAN
>>
>> config CC_NEEDS_LITTLE_ENDIAN
>>           def_bool $(cc-option -mlittle-endian) && CPU_LITTLE_ENDIAN
>>
>> config CC_HAS_FOO
>>          bool
>>          default $(cc-option -mbig-endian -foo) if CC_NEEDS_BIG_ENDIAN
>>          default $(cc-option -mlittle-endian -foo) if CC_NEEDS_LITTLE_ENDIAN
>>          default $(cc-option -foo)
>>
>>
>>
>>
>> Hmm, I think I can implement those somehow.
>> But, I hope we do not have many instances like this...
>>
>>
>> If you know more naive cases, please share your knowledge.
>>
>> Thanks!
>>
>>
>> --
>> Best Regards
>> Masahiro Yamada
>
> Would get pretty bad if a test needs to consider multiple symbols.
> Exponential explosion there...
>
>
> I thought some more about the implementation of dynamic (post-parsing)
> functions to see how bad it would get with the current implementation.
>
> Some background on how things work now:
>
>   1. All expression operands in Kconfig are symbols.
>
>   2. Returning '$ENV' or '$(fn foo)' as a T_WORD during parsing gets
>      you symbols with those strings as names and S_UNKNOWN type (because
>      they act like references to undefined symbols).
>
>   3. For "foo-$(fn foo)", you also get a symbol with that string as its
>      name and S_UNKNOWN type (stored among the SYMBOL_CONST symbols)
>
>   4. Symbols with S_UNKNOWN type get their name as their string value,
>      and the tristate value n.
>
> So, if you do string expansion on the names of symbols with S_UNKNOWN
> type in sym_calc_value(), you're almost there with the current
> implementation, except for the tristate case.
>
> Maybe you could set the tristate value of S_UNKNOWN symbols depending on
> the string value you end up with. Things are getting pretty confusing at
> that point.
>
> Could have something like S_DYNAMIC as well. More Kconfig complexity...
>
> Then there's other complications:
>
>   1. SYMBOL_CONST is no longer constant.
>
>   2. Dependency loop detection needs to consider symbol references
>      within strings.
>
>   3. Dependency loop detection relies on static knowledge of what
>      symbols a symbol depends on. That might get messy for certain
>      expansions, though it might be things you wouldn't do in practice.
>
>   4. Symbols still need to be properly invalidated. It looks like at
>      least menuconfig just does a dumb invalidate-everything whenever
>      the value of a symbol is changed though, so it might not require
>      extra work. (Bit messier in Kconfiglib, which does minimal
>      invalidation to keep scripts fast, but just need to extract a few
>      extra deps there.)
>
>
> It looks like dynamic functions could get quite messy, but might be
> doable if absolutely required. There's probably more devils in the
> details though.
>
> I don't think the static function model precludes switching models later
> btw, when people have more experience.



I really want to start with the static function model
and see if we need the dynamic function implementation.

Here is an idea for the migration path in case
we need to do that in the future.



Currently, every time user input is given,
sym_clear_all_valid() is called.

It is not efficient to blindly re-evaluate expensive $(shell ...)


So, have a list of symbols the function depends on
in its arguments.

For example,

config CC_HAS_SANE_STACKPROTECTOR
        def_bool $(shell $srctree/scripts/gcc-has-stack-protector.sh
$CC $(1), CFLAGS_BASE)


Here the first argument
  $srctree/scripts/gcc-x86-has-stack-protector.sh $CC $(1)

is the shell command.
$(1), $(2), ... will be replaced with the values of symbols (or expressions)
that follow when running the shell command.


The second argument
CFLAGS_BASE
is the dependency symbol (or expression).


CFLAGS_BASE can be dynamically changed like

config CFLAGS_BASE
        string
        default "-m64" if 64BIT
        default "-m32"


When and only when CFLAGS_BASE is updated, the function should be re-calculated.
(This will require efforts to minimize the amount of re-evaluation.)




cc-option will be implemented like follows:

macro cc-option $(shell $CC -Werror $$(1) $(1) -c -x c /dev/null -o
/dev/null, CFLAGS_BASE)



Please notice the difference between $$(1) and $(1).

$(1) is immediately expanded by cc-option macro.

$$(1) is escaped since we want to expand it by $(shell ...), not by
$(cc-option ...)



For example,

  $(cc-option -fstack-protector)

will be expanded to

  $(shell gcc -Werror $(1) -fstack-protector -c -x c /dev/null -o
/dev/null, CFLAGS_BASE)

Since macros are just textual shorthand, so this expansion happens
during the parse phase.



Then, the evaluation phase does the following every time CFLAGS_BASE is updated.

gcc -Werror [value of CFLAGS_BASE] -fstack-protector -c -x c /dev/null
-o /dev/null


This is a new form of expression, so it will be managed in AST tree
with a flag E_SHELL (or E_FUNC) etc.


Not implemented at all.  Just a rough sketch.


-- 
Best Regards
Masahiro Yamada